A Modified Naïve Bayesian-based Spam Filter using Support Vector Machine

Abstract

The ever-growing problem which is threatening the

current mailing system is spam. Spam is nothing but an

unsolicited bulk e-mail frequently sent in a financial nature

which generates the need for creating an anti-spam filter.

Amongst many spam filtering techniques, the most advanced

method "Naïve Bayesian filtering" using the Support Vector

Machine (SVM) have been implemented. Spammers are very

careful about the filtering techniques. For that very reason,

dynamic filtering is needed and the proposed method meets the

demand. The algorithm splits the received email into tokens and

uses Bayes' theorem of probability to calculate the probability of

spam for each token to determine the total spam probability of

the mail. Implementation of SVM instead of corpora is one of the

added features of the algorithm. The most challenging feature

was to take the words as well as whole sentences as input in the

SVM as tokens and feature vectors. The inclusion of sentences in

the dataset training has increased the accuracy of detecting spam

and ham. Natural Language Tool Kit (NLTK) has been used as a

useful language processing tool to tokenize the sentences and

also to understand the meaning of the same types of sentences to

some extent. As a test mail is being compared by word to word

and also sentence to sentence from the training datasets to

determine if the mail is spam or not, it will improve the

performance of the filter. With some simple modifications, the

filter can be used in both server and client end. The efficiency

increases gradually with the increased number of email it

processes.